Advancements in Hydrodynamic Simulations of Urban Water Systems and Wastewater

Management: An Integrative Approach Employing Physics-Guided Machine Learning and

Introduction

Urban water management is an essential public health and environmental stewardship responsibility,
demanding precise planning and real-time response to various challenges such as flooding, supply
distribution, and waste treatment. Hydrodynamic simulations play a crucial role in these management
systems by predicting water flow and behavior under different scenarios. However, traditional simulation
methods are often hampered by computational limitations and model simplifications, leading to less
optimal management decisions. The advent of machine learning offers a promising avenue to overcome
these limitations. Specifically, physics-guided machine learning (PGML) and Deep Learning (DL)
models integrates traditional physics-based models with the adaptive and predictive capabilities of
machine learning algorithms. This research proposes to accelerate hydrodynamic simulations in urban
water management systems by leveraging the strengths of PGML and DL models, thus aiming to enhance
the accuracy, efficiency, and scalability of these critical simulations.

Background

Computational modeling plays a pivotal role in the comprehensive management of urban water systems,
extending across design, monitoring, and operational control phases. Predominantly, deterministic models
grounded in physics, which systematically address Saint Venant's equations in incremental time steps for
each pipeline segment, are employed universally (Abbott and Ionescu, 1967). This particular
methodology, herein referred to as high-fidelity models, thoroughly integrates the intrinsic physical
characteristics of the system, such as pipe diameter and length, facilitating simulations of water levels and
flow dynamics with a high degree of accuracy. This precision renders them highly appealing to industry
professionals. Nonetheless, the requisite prolonged durations for simulation inherent in high-fidelity
models pose significant constraints. Specifically, these extended simulation times notably restrict
stakeholder participation in the preliminary design stage (Leskens et al., 2014), and impede the
comprehensive exploration of the potential impacts stemming from profound uncertainties associated
with future climatic conditions and urban development trajectories (Löwe et al., 2020). Furthermore, the
practical application of high-fidelity models in scenarios necessitating real-time responses, such as in

1
early warning systems and dynamic control mechanisms, is often impractical due to the prohibitive
simulation durations required.

To mitigate these challenges, the field of hydrology has seen the emergence of what are termed lower-
fidelity surrogate models (Razavi et al., 2012). In this vein, an array of conceptual or lumped hydrological
models tailored for urban water systems have been introduced (Kroll et al., 2017; Thrysøe et al., 2019).
These strategies significantly diminish computation times, albeit through either a simplification of the
modeled processes or a reduction in spatiotemporal resolution. Generally, these models focus on
simulating flows while omitting water levels and consider only a small number of nodes within the
pipeline infrastructure. Consequently, the utility of these surrogate models is restricted to specific uses for
example, they may be suitable for analyzing sewer overflows but fall short in flood risk assessments.
Moreover, these models are not easily configured automatically from pipeline data and face challenges in
adapting to changes in the infrastructure of water systems. In contrast, cellular automata models, although
they streamline the mathematical representation of hydrological processes, maintain a detailed level of
model fidelity (Austin et al., 2014). Nevertheless, to ensure precision, these models require the use of
small simulation time steps, similar to those in numerical simulations, which results in limited
enhancements in processing time, typically improving speed by a factor of around five.

The employment of machine learning techniques in the field of hydrology has become increasingly
prevalent. These methods are particularly used in scenarios where the input to the model includes
physical characteristics of systems, enabling the generation of models that have shown to be adaptable
across different catchment areas, as evidenced in studies on rainfall runoff analysis (Kratzert et al., 2019)
and flood forecasting (Bentivoglio et al., 2022). There has been a concerted effort to ensure these models
adhere to principles such as mass conservation (Frame et al., 2022). In both the realms of rainfall-runoff
assessment and hydraulic studies, machine learning has been fused with numerical analysis tools aiming
to streamline complex numerical equations, reveal hitherto unrecognized correlations, or enhance the
accuracy of models. This has involved the utilization of methods including genetic programming
(Danandeh Mehr et al., 2018), neural networks (Hoge et al., 2022), and curve fitting techniques (Jamali et
al., 2021). Although integrating these techniques can enhance the correlation with observational data
while maintaining a degree of physical comprehensibility, the limitations in computational efficiency
mirror those observed with cellular automata methods. For instance, Jamali et al. (2021) reported
negligible reductions in computation times in the context of resolving the two-dimensional shallow water
equations.

Consequently, there exists an unmet need for surrogate methodologies capable of simulating the intricate
hydraulics represented in high-fidelity models of numerous states, yet significantly diminishing the

2
requisite time for simulations. Unlike approaches that rely solely on data, scientific machine learning
aims to amalgamate machine learning with time-honored techniques from scientific computation and
statistics. Scientific machine learning is experiencing swift advancements, particularly in modeling
systems described by differential and partial differential equations, employing both time-discrete (Chen
and Xiu, 2021) and time-continuous methods (Wang et al., 2021). These developments are not only
rooted in foundational ideas (Lagaris et al., 1998) but are also adapting to modern challenges through the
adoption of recently accessible technologies.

The advanced computational capabilities of neural networks are capable of handling models
encompassing a vast array of state variables, a feature particularly beneficial for managing urban water
systems. These systems frequently require the simulation of water levels and flow across an extensive
network of links and nodes, often numbering in the hundreds or thousands. Moreover, scientific machine
learning models have the capacity to discern complex interrelationships between various states, such as
the water levels in adjacent nodes. They also leverage time-series data to enhance simulation precision
through methodologies that harness temporal trends (Ren et al., 2022). For the field of hydraulics, such
attributes might facilitate the use of larger simulation time steps than those currently documented in
existing research. Crucially, scientific machine learning introduces diverse strategies for integrating an
understanding of physical systems into data-driven models. This can be achieved through several means,
including the design of model architectures informed by physics, the application of mathematical
constraints to ensure physical accuracy, or the adoption of loss functions that are informed by physical
principles (Lu et al., 2021)

Research Gaps/ Needs

This proposal delineates pivotal research needs and gaps within urban water management that necessitate
immediate attention, underpinning the integration of Machine Learning (ML) as a cornerstone for
innovation. A prevalent gap is identified in the insufficiency of physics-based constraints within
hydrodynamic simulations leveraging ML, where the prevailing approach has predominantly favored
data-driven predictions. Such methods, devoid of integrating fundamental physical principles, are
susceptible to inaccuracies under unrepresented training conditions. Furthermore, the absence of a
systematic framework for evaluating and enhancing the efficiency of PGML and DL models in simulating
complex urban hydrodynamic processes signifies a considerable gap in the current research landscape.
This highlights the need for standardized methodologies that proficiently incorporate physical constraints
into ML algorithms, subsequently allowing for a quantitative assessment of the impacts on model
performance. Additionally, the adaptability and scalability of PGML and DL applications within varied

3
urban settings, a critical aspect given the diversity of urban water infrastructures, usage patterns, and
management goals remain underexplored, signifying an urgent need for research advancement.

The novelty of this proposal is manifest in its multidimensional approach. Firstly, it aims to pioneer the
development of a proof of concept for simulating urban water system hydraulics through scientific ML,
creating fast surrogate models renowned for their spatial detail and comprehensive consideration of
hydraulic variables such as flows, levels, and surcharges thereby differentiating from existing lumped-
parameter models. This approach, unprecedented in its application, epitomizes a groundbreaking step in
hydrodynamic simulation. Additionally, the proposal advocates for the utilization of generalized residue
networks, augmenting this novel strategy by embedding inductive biases reflective of physical intuitions
concerning the hydraulic behavior within urban pipe systems, inspired by the work of Chen and Xiu
(2021) and recommendations from Garzon et al. (2022). This method transcends traditional data-driven
models by ensuring the alignment of ML architectures with the intrinsic physics of urban water flows.
Furthermore, the application of these advanced techniques in realistic settings characterized by dynamic
rainfall patterns and complex pipe flow dynamics not only showcases the potential of these methods but
also critically identifies existing bottlenecks and pitfalls, thereby enhancing the practical applicability of
the research. Lastly, this proposal commits to delineating further research directions, firmly grounded on
the findings and established methodologies, to perpetually refine and advance the integration of ML in
urban water management.

In essence, the novelty of this proposal lies in its comprehensive, innovative approach to bridging the
current gaps in urban water management systems through the integration of state-of-the-art ML
technologies, promising a transformative impact on the field.

Data needs

In light of the ambitious goals set forth by this proposal, the success of integrating PGML and DL into
urban water management systems hinges critically on the collection, availability, and quality of relevant
data. The data needs for this project are multifaceted, reflecting the complexity and variability of urban
water systems as well as the advanced computational techniques employed. Understanding and outlining
these needs is essential for the project’s foundation, as they will guide the development, testing, and
optimization of the proposed models. Here are the primary data needs for this research:

Hydrodynamic Data: High-resolution spatial and temporal data on water flow, levels, and surcharges
within urban environments are crucial. This includes both historical records and real-time data capturing
the dynamics of urban water systems under varying conditions.

4
Infrastructure Information: Detailed data on the physical characteristics of urban water infrastructure,
including pipe dimensions, materials, ages, and conditions, are essential for developing accurate models
and simulations.

Environmental and Climatic Data: Comprehensive environmental data including precipitation patterns,
temperature fluctuations, and other relevant climatic variables are necessary to accurately simulate the
environmental conditions impacting urban water systems.

Water Usage Patterns: Data on water usage within urban settings, including consumption rates, peak
usage times, and sources of significant demand, are critical for understanding and predicting the pressures
on urban water systems.

Geospatial Data: High-quality, detailed geospatial data are required for mapping the complex interactions
and dependencies within urban water systems and their surrounding environments.

Research Questions

This proposal poses several research questions to guide the investigation:

1. How can PGML and DL models be designed to accurately and efficiently simulate hydrodynamic
processes in urban water systems?

2. What role do physical constraints play in enhancing the predictive performance and reliability of ML
models in these simulations?

3. How can PGML and DL models be optimized for different urban environments, considering variability
in infrastructure, climate, and water use patterns?

4. What are the computational savings achieved by integrating PGML and DL in hydrodynamic
simulations compared to traditional methods?

5. How can the development of a proof of concept using scientific ML with generalized residue networks,
tailored for simulating urban water system hydraulics, revolutionize our understanding and capability to
manage urban water systems under dynamic conditions?

Work Packages for 4 Years Research Chapters

Each WP corresponds to a year and is focused on achieving specific milestones that build toward the final
objective.

Year 1: Foundations and Framework Development

5
WP1: Comprehensive Literature Review and Theoretical Framework

- Conduct an exhaustive review of existing literature covering hydrodynamic simulations, PGML,

WP2: Preliminary Design of Integration Models

- Initiate the design of PGML and DL models specifically tailored for hydrodynamic simulation
tasks, considering the constraints and requirements identified in WP1.
- Establish partnerships with urban water management stakeholders for real-world data acquisition
and model validation.

Year 2: Model Development and Preliminary Testing

WP3: Development of PGML and DL Models

- Develop a series of PGML and DL models, incorporating physics-based constraints and urban
water management principles.
- Implement models in a controlled simulation environment using synthetic or historically available
data sets.

WP4: Preliminary Validation and Refinement

- Perform preliminary testing of developed models to evaluate predictive accuracy, computational

Year 3: Advanced Testing, Optimization, and Evaluation Framework

WP5: Advanced Model Testing and Optimization

- Enhance model robustness and accuracy through advanced testing, including more complex and
dynamic urban water scenarios.
- Optimize models for computational efficiency and adaptability to diverse urban environments.

WP6: Development of an Evaluation Framework

6
 - Develop a comprehensive evaluation framework that quantifies model performance, efficiency,
and scalability.
- Begin integrating models with real-world urban water management systems for live testing and
validation.

Year 4: Real-World Implementation and Scalability Testing

WP7: Implementation and Final Validation**

- Implement PGML and DL models in partnership with urban water management entities for real-
world operation.
- Conduct extensive validation to compare model performance against traditional hydrodynamic
simulations and real-world outcomes.

WP8: Scalability Testing and Framework Refinement

- Test the scalability of the PGML and DL models across different urban configurations, water
management systems, and climates.
- Refine the evaluation framework based on scalability testing, ensuring it supports model
adaptation and customization for global urban settings.

WP9: Documentation and Dissemination

- Document all findings, methodologies, and best practices developed throughout the project.
- Prepare detailed reports and academic papers (minimum 3) summarizing the research outcomes.
- Host workshops and seminars with stakeholders and the academic community to share
knowledge and encourage the adoption of PGML and DL in urban water management systems.

Conclusion

This comprehensive research proposal aims to revolutionize urban water management systems by
integrating PGML and DL with traditional hydrodynamic simulations. Through a meticulously structured
four-year plan, I anticipate addressing the pivotal research needs and gaps identified, notably the lack of
integrated, adaptable, and scalable models that can efficiently simulate complex hydrodynamic processes
in urban environments. The research outlined in the work packages presents a progressive path from
foundational literature review and theoretical framework development through to the real-world
implementation and scalability testing of the PGML and DL models. By systematically advancing
through stages of development, testing, optimization, and application, this research aims not only to
enhance the accuracy and efficiency of urban water management systems but also to significantly reduce

7
the computational costs associated with traditional simulation methods. A key anticipated outcome of this
project is the establishment of a robust evaluation framework that quantitatively and qualitatively assesses
the effectiveness of the proposed integrated models. This framework will not only validate the models
against traditional methods but also ensure their adaptability to varied urban settings worldwide,
highlighting the global applicability and scalability of the research findings. Moreover, this project
promises to fill the critical research gaps by embedding physics-based constraints in ML algorithms,
thereby enhancing the predictive performance and reliability of the models under diverse and dynamic
conditions. The collaborative approach, involving stakeholders from urban water management entities,
ensures that the models developed are grounded in practical, real-world requirements and challenges.
Ultimately, the successful implementation of this research has the potential to set new benchmarks in
urban water management, offering a novel pathway to tackle the increasingly complex challenges posed
by rapid urbanization and climate change. It is envisaged that the findings from this project will not only
contribute significantly to the academic body of knowledge but also provide practical tools and
methodologies that can be adopted and adapted by urban water managers globally. In doing so, this
research aspires to enhance the resilience, sustainability, and efficiency of urban water systems,
safeguarding public health and environmental quality for future generations.

References

Abbott, M.B., Ionescu, F., 1967. On the numerical computation of nearly horizontal flows. J. Hydraul.
Res. 5, 97–117. https://doi.org/10.1080/00221686709500195.

Austin, R.J., Chen, A.S., Savi´c, D.A., Djordjevi´c, S., 2014. Quick and accurate cellular automata sewer
simulator. J. Hydroinform. 16, 1359–1374. https://doi.org/ 10.2166/hydro.2014.070.

Bentivoglio, R., Isufi, E., Jonkman, S.N., Taormina, R., 2022. Deep learning methods for flood mapping:
a review of existing applications and future research directions. Hydrol. Earth Syst. Sci. Discuss.
https://doi.org/10.5194/hess-2022-83.

Chen, Z., Xiu, D., 2021. On generalized residue network for deep learning of unknown dynamical
systems. J. Comput. Phys. 438, 110362. https://doi.org/10.1016/j. jcp.2021.110362.

Danandeh Mehr, A., Nourani, V., Kahya, E., Hrnjica, B., Sattar, A.M.A., Yaseen, Z.M., 2018. Genetic
programming in water resources engineering: a state-of-the-art review. J. Hydrol. 566, 643–667.
https://doi.org/10.1016/j.jhydrol.2018.09.043.

8
Frame, J., Kratzert, F., Klotz, D., Gauch, M., Shelev, G., Gilon, O., Qualls, L.M., Gupta, H. V, Nearing,
G.S., 2022. Deep learning rainfall-runoff predictions of extreme events. Hydrol. Earth Syst. Sci. 26,
3377–3392. https://doi.org/10.5194/hess-26-3377- 2022.

Garzon, ´ A., Kapelan, Z., Langeveld, J., Taormina, R., 2022. Machine learning-based surrogate
modelling for Urban Water Networks: review and future research R. Palmitessa et al. Water Research 223
(2022) 118972 13 directions. Water Resour. Res., e2021WR031808. https://doi.org/10.1029/
2021WR031808.

Hoge, M., Scheidegger, A., Baity-Jesi, M., Albert, C., Fenicia, F., 2022. Improving hydrologic models for
predictions and process understanding using Neural ODEs. Hydrol. Earth Syst. Sci. Discuss.
https://doi.org/10.5194/hess-2022-56.

Jamali, B., Haghighat, E., Ignjatovic, A., Leit˜ ao, J.P., Deletic, A., 2021. Machine learning for
accelerating 2D flood models: Potential and challenges. Hydrol. Process. 35, 1–14.
https://doi.org/10.1002/hyp.14064.

Kratzert, F., Klotz, D., Herrnegger, M., Sampson, A.K., Hochreiter, S., Nearing, G.S., 2019. Toward
improved predictions in ungauged basins: exploiting the power of machine learning. Water Resour. Res.
55, 11344–11354. https://doi.org/10.1029/ 2019WR026065.

Kroll, S., Wambecq, T., Weemaes, M., Van Impe, J., Willems, P., 2017. Semi-automated buildup and
calibration of conceptual sewer models. Environ. Model. Softw. 93, 344–355.
https://doi.org/10.1016/j.envsoft.2017.02.030.

Lagaris, I.E., Likas, A., Fotiadis, D.I., 1998. Artificial neural networks for solving ordinary and partial
differential equations. IEEE Trans. Neural Netw. 9, 987–1000. https:// doi.org/10.1109/72.712178.

Leskens, J.G., Brugnach, M., Hoekstra, A.Y., Schuurmans, W., 2014. Why are decisions in flood disaster
management so poorly supported by information from flood models? Environ. Model. Softw. 53, 53–61.
https://doi.org/10.1016/j.envsoft.2013.11.003.

Lowe, R., Mair, M., Pedersen, A.N., Kleidorfer, M., Rauch, W., Arnbjerg-Nielsen, K., 2020. Impacts of
urban development on urban water management – limits of predictability. Comput. Environ. Urban Syst.
84, 101546.https://doi.org/10.1016/j. compenvurbsys.2020.101546.

Lu, L., Jin, P., Pang, G., Zhang, Z., Karniadakis, G.E., 2021. Learning nonlinear operators via DeepONet
based on the universal approximation theorem of operators. Nat. Mach. Intell. 3
https://doi.org/10.1038/s42256-021-00302-5.

9
Razavi, S., Tolson, B.A., Burn, D.H., 2012. Review of surrogate modeling in water resources. Water
Resour. Res. 48 https://doi.org/10.1029/2011WR011527.

Ren, P., Rao, C., Liu, Y., Wang, J., Sun, H., 2022. PhyCRNet: physics-informed convolutional-recurrent
network for solving spatiotemporal PDEs. Comput. Methods Appl. Mech. Eng. 389, 114399
https://doi.org/10.1016/j.cma.2021.114399.

Thrysøe, C., Arnbjerg-Nielsen, K., Borup, M., 2019. Identifying fit-for-purpose lumped surrogate models
for large urban drainage systems using GLUE. J. Hydrol. 568, 517–533.
https://doi.org/10.1016/j.jhydrol.2018.11.005.

Wang, S., Wang, H., Perdikaris, P., 2021. Learning the solution operator of parametric partial differential
equations with physics-informed DeepOnets. Sci. Adv. 7, eabi8605.
https://doi.org/10.1126/sciadv.abi8605.

10